Battery life self-adaptive calibration method oriented to cloud-edge collaboration

Abstract

The invention discloses a self-adaptive battery life calibration method for cloud-side cooperation, and belongs to the crossing field of an energy storage system and cloud-side cooperation calculation. According to the invention, a cloud-edge double-layer collaborative framework is provided; an edge end estimates the health state and the residual life of a battery in real time through a recursive least square extended Kalman filtering model; the error observer calculates a prediction error based on a sliding window, and a dynamic threshold triggers an uploading mechanism; the edge end adopts an auto-encoder to compress original time sequence features into abstract vectors, and the abstract vectors and error statistics are uploaded together; the cloud performs incremental learning by using a deep sequential network, and only finely adjusts tail level parameters of which the gradient sensitivity exceeds a threshold value to generate a correction value; and the correction value is compressed and issued to an edge end, local model parameters are updated through weighted fusion, and a covariance matrix is adjusted. The method realizes high-precision life prediction and dynamic calibration, remarkably reduces the communication load, and is suitable for electric vehicles, power grid energy storage and other scenes.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of energy storage systems and cloud-edge collaborative computing, and particularly relates to a battery life adaptive calibration method for cloud-edge collaboration. BACKGROUND

[0002] The current battery life prediction technology mainly adopts two methods: a simplified model based on an edge battery management system (BMS) and a global model based on a cloud. In practical application scenarios such as electric vehicles and grid energy storage, the traditional edge model is limited by the computing power and storage resources of embedded devices, and can only deploy simplified algorithms with low computational complexity, resulting in insufficient generalization ability for new working conditions. Although the cloud model can continuously update the global model using massive data, it is restricted by network delay and bandwidth and cannot meet the millisecond-level real-time response requirement of battery management.

[0003] The existing technology has the following inherent defects:

[0004] Firstly, the prediction model trained by relying on offline experimental data will produce significant model drift phenomena in actual operation due to factors such as changes in environmental temperature, fluctuations in charging and discharging rates, and differences in cell manufacturing, resulting in continuous accumulation of errors in the estimation of state of health and remaining life. Secondly, the edge and cloud lack a dynamic collaboration mechanism and cannot perform closed-loop correction according to real-time working condition changes. This leads to inaccurate power jump warning in the electric vehicle range prediction scenario and overcharging / overdischarging safety risks in the grid energy storage frequency regulation scenario. Attempts to improve the above defects face double difficulties: if the complexity of the edge model is simply increased, it will exceed the real-time computing capacity of the microcontroller; if the cloud full model is relied on for updating, the communication load will increase dramatically and the control timeliness cannot be guaranteed.

[0005] Therefore, there is an urgent need for an adaptive life calibration mechanism that realizes cloud-edge dynamic collaboration and real-time suppression of model drift under resource-limited conditions. SUMMARY

[0006] To solve the above technical problems, the application provides a battery life adaptive calibration method for cloud-edge collaboration to solve the problems existing in the prior art.

[0007] In the first aspect, to achieve the above object, the application provides a battery life adaptive calibration method for cloud-edge collaboration, comprising the following steps:

[0008] The edge, cloud and collaboration module execute:

[0009] The edge estimates the battery state of health and remaining life in real time based on the recursive least squares extended Kalman filter model;

[0010] The error observer calculates the prediction error within a sliding window, and triggers uploading when the error exceeds a dynamic threshold;

[0011] The edge end compresses the original time sequence feature into a summary vector through the autoencoder, and uploads the summary vector and the error statistics to the cloud end together;

[0012] The cloud end adopts a deep time sequence network model, and fine-tunes the tail level parameters of the network to generate a correction amount in an incremental learning manner according to the received summary vector and error statistics;

[0013] The cloud end compresses and issues the correction amount to the edge end;

[0014] The edge end fuses the correction amount to update the local model parameters, and adjusts the covariance matrix of the extended Kalman filter;

[0015] The dynamic threshold is calculated based on historical error quantiles, and the incremental learning only updates the network level whose gradient sensitivity exceeds the threshold.

[0016] Optionally, the process of estimating in real time by the edge end based on the recursive least square extended Kalman filter model comprises:

[0017] The state vector is defined to include the health state, internal resistance and polarization capacitance;

[0018] The covariance noise covariance matrix is adaptively adjusted through a covariance matching strategy;

[0019] The initialization adopts one of the cold start modes of the first charge calibration, the factory template issuance or the first frame of inertia fusion.

[0020] Optionally, the process of calculating the prediction error by the error observer comprises:

[0021] The length of the sliding window is set to 1 hour;

[0022] The health state reference value is obtained by full capacity method measurement or cloud end historical model inference;

[0023] The dynamic threshold takes the 95th percentile value of the error in the last 24 hours.

[0024] Optionally, the structure of the autoencoder comprises:

[0025] A three-layer fully connected network architecture;

[0026] The number of input layer nodes is 256, the number of hidden layer nodes is 64, and the number of output layer nodes is the dimension of the summary vector;

[0027] The ReLU activation function is adopted.

[0028] Optionally, the process of incremental learning comprises:

[0029] A network structure combining TCN and Transformer is adopted;

[0030] Only the last L levels of the network are fine-tuned, and L is dynamically selected according to gradient sensitivity;

[0031] The loss function comprises an L1 norm of prediction error and an L2 regularization term of parameter change.

[0032] Optionally, the process of fusing the edge end correction amount comprises:

[0033] The local model weight is updated according to a weighted formula, and the weight coefficient is adaptively adjusted according to the current error;

[0034] An attenuation factor is applied to the covariance matrix of the extended Kalman filter, and the attenuation factor is negatively correlated with the error;

[0035] When the network is interrupted, the cached parameters or the safety mode are enabled according to the interruption duration.

[0036] In a second aspect, the application further provides a battery life adaptive calibration system for cloud edge collaboration, for implementing a battery life adaptive calibration method for cloud edge collaboration, the system comprising:

[0037] A real-time estimation module is deployed at the edge end, for real-time estimation of the battery health state and remaining life based on a recursive least squares extended Kalman filter model;

[0038] An error monitoring module is used to calculate the prediction error in a sliding window, and to trigger uploading when the error exceeds a dynamic threshold;

[0039] A feature processing module is deployed at the edge end, for compressing original time series features into summary vectors through an autoencoder, and uploading the summary vectors and error statistics to the cloud together;

[0040] A cloud learning module is used to adopt a deep time series network model, to fine-tune the tail level parameters of the network in an incremental learning manner according to the received summary vectors and error statistics, and to generate a correction amount;

[0041] A communication compression module is deployed at the cloud end, for compressing and then distributing the correction amount to the edge end;

[0042] An adaptive calibration module is deployed at the edge end, for fusing the correction amount to update the local model parameters, and adjusting the covariance matrix of the extended Kalman filter;

[0043] The dynamic threshold is calculated based on historical error quantiles, and the incremental learning only updates the network levels whose gradient sensitivity exceeds the threshold.

[0044] In a third aspect, the application further provides a computer terminal device, comprising:

[0045] one or more processors;

[0046] a memory coupled with the processors thereof, storing one or more programs;

[0047] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the battery life adaptive calibration method facing cloud edge collaboration in the first aspect.

[0048] In a fourth aspect, the present application further provides a computer readable storage medium, having stored thereon a computer program, which, when executed by a processor, implements the steps of the battery life adaptive calibration method facing cloud edge collaboration in the first aspect.

[0049] In a fifth aspect, the present application further provides a computer program product, comprising a computer program, which, when executed by a processor, implements the steps of the battery life adaptive calibration method facing cloud edge collaboration in the first aspect.

[0050] Compared with the prior art, the present application has the following advantages and technical effects:

[0051] The battery life adaptive calibration method facing cloud edge collaboration provided by the present application effectively suppresses the model drift problem in battery life prediction through the double-layer model architecture of cloud edge collaboration, combined with the error observer and the adaptive incremental learning mechanism. The edge end adopts a lightweight recursive filtering model to realize millisecond-level real-time estimation, and the cloud end performs global modeling through a deep time series network. When the sliding window error exceeds the dynamic threshold, a feature compression upload mechanism is triggered, and only the abstract vector and error statistics are transmitted. The cloud end performs incremental fine-tuning on the gradient sensitive level to generate a correction amount, which is compressed and sent to the edge end. The edge end fuses the correction amount to update the local parameters and adjust the covariance matrix, forming a closed-loop calibration. This method significantly reduces the communication load while ensuring prediction accuracy, and ensures system robustness through a hierarchical disaster recovery mechanism. BRIEF DESCRIPTION OF DRAWINGS

[0052] The accompanying drawings, which form a part of the present application, are intended to provide further understanding of the present application, and the illustrative embodiments thereof and their descriptions serve to explain the present application. In the drawings:

[0053] Figure 1 is a system architecture diagram of an embodiment of the present application;

[0054] Figure 2 is an RLS-EKF algorithm flowchart of an embodiment of the present application;

[0055] Figure 3A TCN+Transformer network structure diagram of an embodiment of the present application;

[0056] Figure 4 An adaptive calibration timing diagram of an embodiment of the present application. DETAILED DESCRIPTION

[0057] It should be noted that the embodiments in the present application and the features in the embodiments can be combined with each other without conflict. The present application will be described in detail below with reference to the accompanying drawings and in combination with embodiments.

[0058] It should be noted that the steps shown in the flowchart of the accompanying drawings can be executed in a computer system such as a group of computer executable instructions, and although the logical order is shown in the flowchart, in some cases, the steps shown or described herein can be executed in an order different from that shown herein.

[0059] The present application is a method for real-time adaptive calibration of battery remaining useful life (RUL) and state of health (SoH) by interacting with cloud large models and edge light models. The method can be applied to electric vehicles, power grid energy storage, electric tools and consumer electronics, etc. to realize high-precision, low-delay battery life prediction and dynamic calibration. The present application proposes a "cloud-edge double-layer model + error observer + adaptive incremental learning" framework, the goal is:

[0060] Real-time: the edge outputs SoH / RUL within ≤10ms;

[0061] Adaptive: if the model drift error e continuously exceeds the threshold value ε, automatically trigger cloud fine-tuning and incremental delivery;

[0062] Bandwidth friendly: only upload a small amount of summary features and error indicators, communication load ≤50kB / day / car;

[0063] High precision: long-term SoH error ≤2%, RUL prediction relative error ≤5%.

[0064] Embodiment one

[0065] As shown in Figure 1 , the present embodiment provides a battery life adaptive calibration method for cloud-edge collaboration, comprising:

[0066] executed by the edge, the cloud and the collaboration module:

[0067] The edge estimates the battery health state and remaining life in real time based on the recursive least squares extended Kalman filter model;

[0068] The error observer calculates the prediction error in the sliding window, and triggers uploading when the error exceeds the dynamic threshold;

[0069] The edge end compresses the original time sequence features into a summary vector through a self-encoder, and uploads the summary vector and error statistics to the cloud;

[0070] The cloud adopts a deep time sequence network model, and according to the received summary vector and error statistics, fine-tunes the tail level parameters of the network in an incremental learning manner to generate a correction quantity;

[0071] The cloud compresses and issues the correction quantity to the edge end;

[0072] The edge end fuses the correction quantity to update the local model parameters, and adjusts the covariance matrix of the extended Kalman filter;

[0073] The dynamic threshold is calculated based on historical error quantile, and the incremental learning only updates the network level whose gradient sensitivity exceeds the threshold.

[0074] The method is composed of five cooperative modules:

[0075] Edge end lightweight model Real-time estimation of SoH / RUL based on recursive least squares extended Kalman filter (RLS-EKF);

[0076] System calibration for first frame data processing or first deployment scenario:

[0077] First charge calibration: the system collects complete voltage and current sequences during the first full charge process (first reaching 100% SoC and resting ≥30 minutes) to estimate the initial capacity, internal resistance and other model states; and fixes the initial SoH as 100%, and sets the initial covariance matrix P0 as a weak confidence diagonal matrix.

[0078] Factory template issuance: for devices that cannot collect complete charging data (such as static state before factory), the edge end can obtain the initialization parameter template matched by battery model, factory batch and use environment from the cloud, to realize template-based cold start.

[0079] First frame observation inertia fusion: if the edge device is first connected but not in the stable charging phase, only partial observations (such as voltage) are collected, the system will temporarily suspend state update, collect continuous data segments for mean / gradient calculation, and then perform first EKF update.

[0080] In the subsequent fusion mechanism, the initial error is gradually corrected by the error observer and the cloud calibration module to ensure long-term convergence and stability of the system estimation error.

[0081] Error observer ε: calculate the prediction residual in the sliding window and judge whether it is out of bounds;

[0082] Feature compression upload: only upload k-dimensional summary vector f and error statistics e;

[0083] Cloud large model Adopt deep time sequence network (TCN+Transformer) full volume training, and fine-tune in incremental learning mode;

[0084] Cloud large model M_cloud: input data format and standardization processing method

[0085] The input data format is shown in Table 1, which represents the window tensor X t Key parameters:

[0086] Table 1

[0087]

[0088]

[0089] The composition of feature dimension d is shown in Table 2:

[0090] Table 2

[0091]

[0092] The above is the commonly used configuration; Parameters such as Δt and L can be flexibly adjusted in config.yaml, and the 6 in "original measurement (6)" in Table 2 refers to the number of features, and the derived features and static features are the same as above, with 5 and 2 features respectively.

[0093] As an embodiment in this embodiment, the process of real-time estimation of the edge based on the recursive least square extended Kalman filter model includes:

[0094] Define the state vector to include the health state, internal resistance and polarization capacitance;

[0095] Adaptively adjust the process noise covariance matrix through the covariance matching strategy;

[0096] Initialization adopts one of the cold start modes of first charge calibration, factory template issuance or first frame inertia fusion.

[0097] Specifically, the standardization and preprocessing process includes:

[0098] Offline statistics: calculate the mean μ of each continuous feature on the training set i And standard deviation σ i If the feature has natural upper and lower limits (such as SOC), use:

[0099]

[0100] Zero mean unit variance scaling (Z score):

[0101]

[0102] μ i ,σ i Saved together with model parameters, shared during inference.

[0103] Winsorize: if then truncated to ±5, with abnormal rate recorded.

[0104] Missing value completion: linear interpolation for continuous missing values ≤3 steps; 3 steps filled with the mean of the last valid window, and set to 1 in the mask channel.

[0105] Online incremental update: Welford algorithm exponential smoothing update during fine-tuning phase:

[0106]

[0107] Where α = 0.001.

[0108] Batch packing: normalized window Input TCN→Transformer main body; static features are broadcasted and spliced to each time step.

[0109] Adaptive calibrator Fusion of cloud correction amount Δθ and edge model parameters to realize online recursive calibration.

[0110] The interface agreement of the cloud edge collaboration is shown in Table 3:

[0111] Table 3

[0112] Flow Direction Content Trigger condition / Explanation Upload Edge→Cloud Abstract tensor Z, anomaly mask Newest μ i ,σ i normalization]]> Downlink Cloud→Edge correction amount Δθ, updated μ i ,σ i ]]> when KL divergence > 10 -3 ]]

[0113] As an embodiment in the embodiment, the process of the error observer calculating the prediction error comprises:

[0114] The sliding window length is set to 1 hour;

[0115] The health state reference value is obtained by full charge capacity method or cloud historical model inference;

[0116] The dynamic threshold is the 95th percentile value of the error in the last 24 hours.

[0117] Specifically, as shown in Figure 2 , the edge RLS-EKF estimation comprises:

[0118] State space:

[0119] The state vector is defined as:

[0120]

[0121] Measurement equation:

[0122] U k = OCV(SoC k )- I k R 0,k -V 1,k +v k ,

[0123] Process equation:

[0124] x k+1 = Ax k +w k , A = I + AtA,

[0125] where A is the diagonal matrix of decay rates.

[0126] RLS-EKF update:

[0127]

[0128] P k|k = (I - K k H k )P k|k-1 .

[0129] And use recursive least squares to adaptively adjust the process noise covariance Q.

[0130] Adopt the "covariance matching-RLS" strategy to update after each observation:

[0131]

[0132] where is the innovation (residual), and λ ∈ (0, 1) is the forgetting factor (typical value λ = 0.99). When any diagonal element <q k in Q min is truncated to q min , prevent numerical underflow.

[0133] As an embodiment in the embodiment, the structure of the autoencoder comprises:

[0134] a three-layer fully connected network architecture;

[0135] the number of nodes in the input layer is 256, the number of nodes in the hidden layer is 64, and the number of nodes in the output layer is the dimension of the summary vector;

[0136] a ReLU activation function is used.

[0137] Specifically, the error observation and uploading strategy comprises:

[0138] Within a sliding window length W (such as 1h), calculate:

[0139]

[0140] If e SoH > ε SoH or e RUL > ε RUL , then pack and upload:

[0141]

[0142] where s k is the original time-series feature, AE edge is the edge autoencoder.

[0143] Reference values SoH ref , RUL ref are obtained:

[0144] Full capacity method (measured): Perform a “constant current-constant voltage to full charge” test every τ cap (typically 14d), record:

[0145] C full (k) = ∫I chg dt, SoH ref = C full (k) / C rated .

[0146] Cloud calibration (historical model values): If the site cannot be shut down for a long time, then the latest working conditions are inferred by the cloud TCN + Transformer large model (SoH cloud , RUL cloud ), as the reference sequence issued.

[0147] Sliding window error and threshold setting:

[0148] Sliding window length W = 1h = 3600 / Δt steps, calculate:

[0149]

[0150] Dynamic threshold rule:

[0151]

[0152] where ρ = 0.05 (95% quantile), N hist is the error history of the last 24h. If trigger the upload. When , the formula means “the error of the state of health is greater than the threshold value, and it is reported”; when The formula states that "if the error in Remaining Useful Life exceeds a threshold, it should be reported."

[0153] Uplink Packing Features and Edge Autoencoder:

[0154] When the triggering condition is met, the following will occur:

[0155]

[0156] Uploaded to the cloud, including:

[0157] s k =[V k ,I k ,T k [,...] represents the original time series features;

[0158] AE edge ={φ θ ,ψ θ} is a sequence autoencoder: f = φ θ (S),

[0159] Training mechanism summary:

[0160] Offline pre-training: Experimenting with factory datasets minimize on The initial parameter θ0 is obtained.

[0161] Online fine-tuning: After deployment, with a buffer zone (Data from the most recent M hours) consists of unlabeled samples, executed periodically:

[0162] η = 10 -4 M = 6h.

[0163] Drift detection: If If the sequence exceeds the 99th percentile of the training phase, then f will be uploaded together with the original sequence for retraining in the cloud.

[0164] As one implementation method in this embodiment, the incremental learning process includes:

[0165] A network structure combining TCN and Transformer is adopted;

[0166] Only the parameters of the last L layers of the network are fine-tuned, and L is dynamically selected based on gradient sensitivity.

[0167] The loss function includes the L1 norm of the prediction error and the L2 regularization term for parameter variation.

[0168] Specifically, such as Figure 3As shown, the process of cloud-based deep time series modeling and incremental learning includes:

[0169] The network structure includes:

[0170] Input feature sequence {s t First, local dependencies are extracted using a Temporal Convolutional Network (TCN), then fed into a multi-head self-attention Transformer to obtain the global representation h. t Output SoH / RUL estimate:

[0171]

[0172] Incremental fine-tuning:

[0173] Receive batch uploads (f, e), with loss:

[0174]

[0175] Only the weights of the last L layer are fine-tuned to generate a correction value Δθ. This value is then compressed and sent to the corresponding edge device.

[0176] The logic for selecting the fine-tuning layer L includes:

[0177] Feature stability: The front layer (embedding+TCN) learns general battery physical features, which change slowly with operating conditions; excessive updates will weaken the shared representation between cross-battery data and impair generalization ability.

[0178] Parameter efficiency: Given a total network parameter count of N, only the last L layers (N...) are updated. tail ≈5%–10% N) can make Significantly reduced, lowering edge bandwidth consumption.

[0179] Lessons learned: Comparing "freezing 90%" and "full fine-tuning" on public datasets (NASA B0005, CALCE A123), the former showed better performance on MAE. SoH An increase of less than 0.15% while a decrease in upload volume of approximately 20 times.

[0180] Adjustable strategy: Set sensitivity index If S exists l >S th Then the corresponding layer is added to the "fine-tunable set" to form the adaptive-L scheme.

[0181] The cloud-based fine-tuning and correction distribution process includes:

[0182] Receive batch uploads Then, the cloud only fine-tunes the weights of the last L levels, using the loss function:

[0183]

[0184] Obtain correction amount After 8-bit quantization + Huffman compression, it is issued to the corresponding edge device.

[0185] As an embodiment in this embodiment, the process of fusing the correction amount at the edge includes:

[0186] Update the local model weight according to the weighted formula, and the weight coefficient is adaptively adjusted according to the current error;

[0187] An attenuation factor is applied to the covariance matrix of the extended Kalman filter, and the attenuation factor is negatively correlated with the error;

[0188] When the network is interrupted, the cache parameters or the safety mode are enabled according to the interruption duration.

[0189] Specifically, as shown in Figure 4 The adaptive calibrator includes:

[0190] The edge receives Δθ, and according to:

[0191] θ edge ←(1-λ)θ edge +λΔθ,

[0192] Where λ is the calibration step (0.05-0.2). At the same time, an attenuation coefficient ξ<1 is applied to the covariance matrix of RLS-EKF, and the cloud correction result is quickly fused.

[0193] The edge weight fusion step and the dynamic step λ:

[0194] After the edge device receives Δθ, according to:

[0195] θ edge ←(1-λ)θ edge +λΔθ,

[0196] Where λ∈[0.05,0.2] adopts error-driven adaptive:

[0197]

[0198] λ min =0.05,λ max =0.20,ε=10 -3 .

[0199] RLS-EKF covariance attenuation factor ξ:

[0200] Covariance matrix adjustment P k|k ←ξ k P k|k ,ξ k =ξmin +(ξ max -ξ min )exp(-γe k ),

[0201] ξ min =0.85,ξ max =0.98, γ=40.

[0202] The larger the error ξ k The smaller the value, the greater the weight given to the cloud correction results, thus accelerating the filter convergence speed.

[0203] Fault tolerance and degradation strategies:

[0204] Local cache: Cache the latest (Δθ,λ) k ,ξ k ) to / var / syt / cache / ; if the network interruption is ≥5min, continue executing RLS–EKF+edge-online fine-tuning (learning rate gradually decays to 10). -6 ).

[0205] Tiered downgrade:

[0206] a. Interruption of 1–30 min: Use the last issued (Δθ,λ) k ,ξ k ).

[0207] b. Interruption of 30 min–6 h: Freeze θ edge RLS–EKF maintains a small process noise Q safe .

[0208] c.>6h Interrupt: Enter Safe-Mode and only output SoC / power protection related alarms.

[0209] Reconnection and Synchronization: After the connection is restored, immediately request the latest Δθ. full (≥24h cumulative).

[0210] Sample verification is performed based on the above method:

[0211] Setting scenario: 100 electric vehicles, 3 months of driving data; cloud GPU cluster; edge MCU Cortex-M7.

[0212] The comparison results of the indicators are shown in Table 4:

[0213] Table 4

[0214] Method SoH MAE (%) RUL MAPE (%) Communication (kB / day) Edge-only static model 4.8 11.3 0 Cloud-only offline model 3.2 7.5 300 The method of the present invention 1.9 4.6 48

[0215] The SoH error is kept less than 2% within 3 months, and the communication load is only 16% of that of the pure cloud scheme.

[0216] Based on this, the battery life adaptive calibration method for cloud edge collaboration provided by the embodiment of the application has the following advantages compared with the prior art:

[0217] 1. Cloud edge complementarity: the cloud global model captures complex aging mechanisms, and the edge end quickly responds to local working conditions.

[0218] 2. Closed-loop adaptation: error-triggered upload and incremental delivery, realizing continuous calibration rather than one-time training;

[0219] 3. Low bandwidth and high security: only feature summaries and correction amounts are transmitted to avoid leakage of raw data;

[0220] 4. Hardware friendly: edge algorithm complexity O(n), which can be completed within 5ms on a 240MHz MCU;

[0221] 5. High accuracy: long-term SoH / RUL error is significantly lower than that of traditional single-layer models.

[0222] Embodiment two

[0223] In this embodiment, a computer terminal device is provided, comprising:

[0224] one or more processors;

[0225] a memory coupled to the processor for storing one or more programs;

[0226] When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the above-mentioned battery life adaptive calibration method for cloud edge collaboration.

[0227] In this embodiment, a computer readable storage medium having a computer program stored thereon is also provided, and the computer program is executed by a processor to implement the steps of the above-mentioned battery life adaptive calibration method for cloud edge collaboration.

[0228] In this embodiment, an electronic device is also provided, comprising a memory and a processor, the memory having a computer program stored therein, and the processor is configured to run the computer program to perform the steps of the above-mentioned battery life adaptive calibration method for cloud edge collaboration.

[0229] In this embodiment, a computer program product is also provided, comprising a computer program, which is executed by a processor to implement the steps of the above-mentioned battery life adaptive calibration method for cloud edge collaboration.

[0230] The above program can run in a processor, or can also be stored in a memory (or called a computer readable medium), the computer readable medium includes permanent and non-permanent, removable and non-removable media, which can realize information storage by any method or technology. Information can be computer readable instructions, data structure, program module or other data. Examples of computer storage medium include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disc read-only memory (CD-ROM), digital versatile disc (DVD) or other optical storage, magnetic cassette, magnetic tape disk storage or other magnetic storage device or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0231] These computer programs can also be loaded into a computer or other programmable data processing device, so that a series of operation steps are executed on the computer or other programmable device to generate a computer implemented process, so that the instructions executed on the computer or other programmable device provide a process for implementing the functions specified in the flow Figure 1 One flow or multiple flows and / or blocks Figure 1 One block or multiple blocks, the steps of the functions specified in the flow or the block can be realized by different modules.

[0232] Such a device or system is provided in the embodiment. The system is called a cloud edge collaborative battery life adaptive calibration system, which includes:

[0233] A real-time estimation module is deployed at the edge end, which is used to estimate the battery health state and remaining life in real time based on a recursive least squares extended Kalman filter model;

[0234] An error monitoring module is used to calculate the prediction error in the sliding window, and trigger uploading when the error exceeds the dynamic threshold;

[0235] A feature processing module is deployed at the edge end, which is used to compress the original time series features into summary vectors through an autoencoder, and upload the summary vectors and error statistics to the cloud together;

[0236] A cloud learning module is used to adopt a deep time series network model, and fine tune the network tail level parameters to generate correction according to the received summary vectors and error statistics in an incremental learning manner;

[0237] A communication compression module is deployed in the cloud, which is used to compress and issue the correction to the edge end;

[0238] An adaptive calibration module deployed at the edge end is configured to fuse a correction amount to update a local model parameter and adjust a covariance matrix of an extended Kalman filter.

[0239] The dynamic threshold is calculated based on historical error quantile, and the incremental learning only updates network levels with gradient sensitivity exceeding the threshold.

[0240] As an embodiment in the present embodiment, the real-time estimation module comprises:

[0241] A state definition unit is configured to define a state vector comprising a health state, an internal resistance and a polarization capacitance.

[0242] A noise adjustment unit is configured to adaptively adjust a process noise covariance matrix through a covariance matching strategy.

[0243] An initialization unit is configured to perform one of a cold start mode, such as a first charge calibration, a factory template distribution or a first frame inertia fusion.

[0244] As an embodiment in the present embodiment, the error monitoring module comprises:

[0245] A window configuration unit is configured to set a sliding window with a length of 1 hour.

[0246] A reference value acquisition unit is configured to acquire a health state reference value through a full charge capacity method or a cloud historical model inference.

[0247] A threshold calculation unit is configured to take a 95% quantile value of errors in the last 24 hours as a dynamic threshold.

[0248] As an embodiment in the present embodiment, the autoencoder in the feature processing module comprises:

[0249] An input encoding unit is configured with a 256-node fully connected layer.

[0250] A feature compression unit is configured with a 64-node hidden layer.

[0251] An abstract output unit is configured with an output layer node number equal to the dimension of the abstract vector.

[0252] An activation unit adopts a ReLU activation function to perform a nonlinear transformation.

[0253] As an embodiment in the present embodiment, the cloud learning module comprises:

[0254] A time series feature extraction unit adopts a TCN network structure to extract local dependence.

[0255] A global modeling unit adopts a Transformer network structure to generate a global representation.

[0256] A parameter fine-tuning unit is configured to fine-tune only the last L levels of parameters of the network, where L is dynamically selected according to the gradient sensitivity.

[0257] A loss control unit is configured to use a loss function including a prediction error L1 norm and a parameter change L2 regularization term.

[0258] As an implementation in the embodiment, the adaptive calibration module includes:

[0259] A weight fusion unit is configured to update the local model weights according to a weighted formula, and the weight coefficients are adaptively adjusted according to the current error.

[0260] A covariance attenuation unit is configured to apply an error-negative correlation attenuation factor to the covariance matrix of the extended Kalman filter.

[0261] A disaster recovery processing unit is configured to enable the cache parameters or the safety mode according to the interruption duration when the network is interrupted.

[0262] The system or device is used to realize the functions of the methods in the above embodiments. Each module in the system or device corresponds to each step in the method, and has been described in the method and will not be repeated here.

[0263] Through the above implementation, the problem of battery life adaptive calibration for cloud-edge collaboration in the related art is solved, thereby ensuring the problems in the prior art to be solved.

[0264] The above is only a preferred specific embodiment of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art can easily think of changes or replacements within the technical scope disclosed by the present application, which should be covered within the protection scope of the present application. Therefore, the protection scope of the present application should be subject to the protection scope of the claims.

Claims

1. A battery life adaptive calibration method for cloud edge collaboration, characterized in that, The method comprises the following steps: The edge end, the cloud end and the coordination module perform: The edge end estimates the battery health state and the remaining life in real time based on a recursive least square extended Kalman filter model; An error observer calculates the prediction error in a sliding window, and triggers uploading when the error exceeds a dynamic threshold; The edge end compresses the original time series features into a summary vector through a self-encoder, and uploads the summary vector and the error statistics to the cloud end; The cloud end adopts a deep time series network model, adjusts the tail level parameters of the network in an incremental learning manner according to the received summary vector and error statistics, and generates a correction quantity; The cloud end compresses and issues the correction quantity to the edge end; The edge end fuses the correction quantity to update the local model parameters and adjusts the covariance matrix of the extended Kalman filter. The dynamic threshold is calculated based on historical error quantiles, and the incremental learning only updates the network levels whose gradient sensitivity exceeds the threshold.

2. The method of claim 1, wherein, The process of estimating the battery health state and the remaining life in real time by the edge end based on the recursive least square extended Kalman filter model comprises: Defining a state vector including the health state, internal resistance and polarization capacitance; Adaptively adjusting the covariance matrix of the process noise through a covariance matching strategy; Initializing in a cold start mode selected from one of the first charge calibration, factory template issuance or first frame inertia fusion.

3. The method of claim 1, wherein, The process of calculating the prediction error by the error observer comprises: The length of the sliding window is set to 1 hour; The health state reference value is obtained by full capacity method measurement or cloud end historical model inference; The dynamic threshold takes the 95th quantile value of the error in the last 24 hours.

4. The method of claim 1, wherein, The structure of the self-encoder comprises: A three-layer fully connected network architecture; The number of input layer nodes is 256, the number of hidden layer nodes is 64, and the number of output layer nodes is the dimension of the summary vector; Using a ReLU activation function.

5. The method of claim 1, wherein, The process of incremental learning comprises: Using a network structure combining TCN and Transformer; Only the last L level parameters of the network are adjusted, and L is dynamically selected according to the gradient sensitivity; The loss function includes the L1 norm of the prediction error and the L2 regularization term of the parameter change.

6. The method of claim 1, wherein, The process of the edge end fusing the correction quantity comprises: Updating the local model weight according to a weighted formula, and adaptively adjusting the weight coefficient according to the current error; Applying a decay factor to the covariance matrix of the extended Kalman filter, and the decay factor is negatively correlated with the error; When the network is interrupted, the buffered parameters or the safety mode are enabled according to the interruption time. 7.A cloud-edge collaboration oriented battery life adaptive calibration system, characterized in that, The system comprises: A real-time estimation module deployed on the edge end, for estimating the battery health state and the remaining life in real time based on a recursive least square extended Kalman filter model; An error monitoring module for calculating the prediction error in a sliding window, and triggering uploading when the error exceeds a dynamic threshold; A feature processing module deployed on the edge end, for compressing the original time series features into a summary vector through a self-encoder, and uploading the summary vector and the error statistics to the cloud end; A cloud end learning module for adopting a deep time series network model, adjusting the tail level parameters of the network in an incremental learning manner according to the received summary vector and error statistics, and generating a correction quantity; A communication compression module deployed on the cloud end, for compressing and issuing the correction quantity to the edge end; An adaptive calibration module deployed at the edge end is configured to fuse correction quantity to update local model parameters and adjust a covariance matrix of an extended Kalman filter; The dynamic threshold is calculated based on historical error quantile, and the incremental learning only updates the network level whose gradient sensitivity exceeds the threshold.

8. A computer terminal device, characterized by Comprise: One or more processors; Memory, coupled with the processor, for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the steps of the method of any one of claims 1-6.

9. A computer readable storage medium having stored thereon a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6.

10. A computer program product comprising a computer program, characterized in that, The computer program is executed by the processor to implement the steps of the method of any one of claims 1-6.

Images